Automatic pterygopalatine fossa segmentation and localisation based on DenseASPP

Abstract

Background

Allergic rhinitis constitutes a widespread health concern, with traditional treatments often proving to be painful and ineffective. Acupuncture targeting the pterygopalatine fossa proves effective but is complicated due to the intricate nearby anatomy.

Methods

To enhance the safety and precision in targeting the pterygopalatine fossa, we introduce a deep learning-based model to refine the segmentation of the pterygopalatine fossa. Our model expands the U-Net framework with DenseASPP and integrates an attention mechanism for enhanced precision in the localisation and segmentation of the pterygopalatine fossa.

Results

The model achieves Dice Similarity Coefficient of 93.89% and 95% Hausdorff Distance of 2.53 mm with significant precision. Remarkably, it only uses 1.98 M parameters.

Conclusions

Our deep learning approach yields significant advancements in localising and segmenting the pterygopalatine fossa, providing a reliable basis for guiding pterygopalatine fossa-assisted punctures.